Mission-Critical Forecasting

(Published in Mid-Range ERP magazine November 1997)

Background Note: Although this article was published almost a decade ago, the tenets and concepts expressed herein continue to hold true today. In fact, these beliefs have become all the more important as we participate in an increasingly globalized business environment.

Forecasting is a subject increasingly under scrutiny as corporations look to improve their business processes and profitability. As the business world becomes ever more competitive, many organizations are finding that better forecasting benefits their bottom line directly. This article is an attempt to clarify some concepts of forecasting, and to discuss some of the general issues that must be dealt with when designing a forecasting system and process.

A good place to start is to define the term forecasting, which is often misunderstood. Forecasting is a process of using various tools and techniques to anticipate the amounts and/or values of future sales of products and services. It is not a measurement of a known quantity, but an attempt to combine historically observed patterns with "known" influencing events to come up with a "most educated guess" as to what the future holds. Forecasting is therefore a planning activity, fundamentally different from most measurement activities. It is constantly focused not on what has happened but on what might happen. Whereas someone might set out to become proficient at accounting, and become virtually perfect at keeping debits and credits balanced, in forecasting there is always a degree of uncertainty and error that no amount of skill or training will overcome. The challenge is to set up software systems as well as processes of information flow that allow getting closer to the desired benchmark of accuracy.

Given the difficulties inherent in forecasting, companies have in the past been tempted to neglect this area of activity. They are reminded, however, by some of the best-known consultants, that this cannot be overlooked. According to "The Oliver Wight ABCD Checklist for Operational Excellence", in order for a manufacturing organization to qualify as a "Class A" MRP II operation:

"There is a process for forecasting all anticipated demands with sufficient detail and adequate planning horizon to support business planning, sales and operations planning, and master production scheduling. Forecast accuracy is measured in order to continuously improve the process."

In other words, not only is this activity essential, but it's accuracy is a metric that should be monitored. Although each year brings new gurus and acronyms to the industry, this is one of the basic concepts that has not changed.

Unfortunately, while the importance of forecasting has been recognized, there has been a shortage of information and training as to how an effective forecasting process should be designed. As already mentioned, it involves not only software systems, but also processes where people exchange information. As such it is not an exact science, but to a certain extent an art as well. We have to accept a basic premise: The forecasts are always wrong! (If you were 100% accurate, you would never be wrong). The focus becomes "Being less wrong, as consistently as possible". The focus is also on on-going review and corrective action to do better next time.

The next issue is to discuss who, meaning what kind of organizations, need to forecast. Traditionally, forecasting was used in make-to-stock environments, where companies would maintain inventories of regularly sold products, since they were never quite sure what the demand would be, and would therefore maintain cycle stock plus some safety stock. This could apply to products manufactured or distributed. In a make-to-order environment, companies traditionally thought it was unnecessary to forecast. When an order came in from a customer, that order was executed because it in effect was the forecast. This is changing rapidly however, as customers have become far more demanding, and expect very short turn-around times between order and delivery, thus forcing make-to-order companies to start anticipating demands before they become known.

Once an organization has decided to start forecasting, a crucial decision is the level of detail to choose. Data is usually available from transaction detail right up to corporate totals. Transactions are usually not forecastable, it is best to work with summarized snapshots, such as weekly, 13-period or monthly data. Data can also be looked at at various levels of corporate reporting hierarchies (Company vs. Division vs. Product family vs. SKU) , or segmented geographically (Country vs. Region vs. Sales Territory vs. Customer). The impacts of this decision are very far-reaching. If you forecast at a high level of summarization, you will likely achieve very accurate results, but the forecast will be not very useful to running the enterprise. If you forecast at too low a level of detail, you will be overburdening your users, and the accuracy will be poor. The rule of thumb is: As the level of detail increases, the workload increases dramatically, while accuracy may increase to a point and then will drop off. In other words, it is crucial to find an optimal level of detail / summarization to forecast at, where your forecasters have a reasonable workload, and where they can still add value to the forecasts. Bear in mind that for best results, the forecast is created by a combination of machine-generated baseline forecasts combined with user intervention. A key ratio that should be looked at is: Forecasted Items/Forecasters.

Another important driver of the process design will be to determine who the internal customers of the forecasts will be. Is the forecast primarily to feed MRP / Master Production Scheduling or is it to help Marketing with their planning? Is it Sales-driven and also used to set Sales quotas and budgets, or is the priority to allow Finance a pro-forma look at revenues derived from product sales? Does Senior Management look at the forecast as a "wished-for" number or will it be used to drive Distribution Center replenishment ? Is the corporate goal a true "1-number forecast" as per the MRP II concept? Will your external customers need to access the forecasts via Internet, or are the forecasts strictly for in-house consumption? The answers to these types of questions go a long way to determining how to set up a process that will serve the needs of all the recipients of the forecast numbers.

The next major decision involves the tools used to generate the forecasts. Organizations using mainframe or legacy forecasting tools are under pressure from users to move to forecasting applications using PC and Windows technology, not only to achieve better integration with other applications, but also to improve ease-of-use with GUI interfaces, ensure Year 2000 compliance and Internet capability. Sales-driven environments are faced with the challenge of improving forecasting in the context of Sales Force Automation.

Defining who should have the responsibility of forecasting can also be a difficult and politically charged decision. If the decision to forecast is driven by a need to feed production scheduling or warehouse replenishment, this will often mean ownership of the process in logistics. An organizations defined along business unit lines or using a customer service team approach will often have forecasting done by Marketing managers or Customer Service managers as one of many functions. Other more centrally organized companies may employ full-time demand planners or forecasting teams that do nothing else.

The methodology of forecasting used to be exclusively in the domain of statistical techniques, which would utilize various mathematical models to capture trend and seasonality patterns from historical data, and extrapolate these into the future. Measures of accuracy were also traditional mathematical metrics such as R-squared, or Mean Absolute Deviation (MAD). This type of machine-driven baseline forecast continues to provide many benefits, due to the automation and time-savings it entails. It worked well in the past for very stable businesses, but is missing a required dimension for the majority of today¹s enterprises.

The additional component needed is that of marketing events or promotions determined from outside the company. These events could range from competitive product launches, government regulatory intervention, pricing and promotional events determined by customers, and more. Any knowledge of such events is usually partial and at best an educated guess. However, most Sales or Marketing people usually have some sense of such upcoming events. This understanding may be as vague as a "gut-feel", or be as concrete as a promise by a Customer to deliver a certain volume of sales on a promotion. The expectation may change frequently as the event gets closer, the timing and magnitude may vary as well, reinforcing our statement that this is more an art than a science. However, this judgmental knowledge of influencing events is crucial data that needs to be incorporated into the forecasts, and can add tremendous value to it. These expected events become documented assumptions underlying the forecast. The requirement of a forecasting system is therefore to allow easy mechanisms for integrating such knowledge, and overlaying it onto a mechanically produced baseline forecast. Measurements of accuracy also have to change along with this methodology, and focus more on variance of actual from forecast than traditional statistical metrics.

Another important aspect to the design a forecasting system and process relates to the sources of data. In the case of creating the baseline machine-generated forecast, the source of data is usually the host system or enterprise system, whether mainframe, mid-range or network-based. Historical data is usually collected from Order Entry/Sales tables, from where it can be downloaded to PC level. Ideally such an interface uses non-proprietary technology such as ODBC. Finished forecasts are most often uploaded into Master Production Scheduling, or in some cases into other planning and simulation systems, reporting systems or even Data Warehouses. The source of Market Event data is usually more difficult, as many companies do not formally collect such information, and would rely on a forecasting system to provide such a mechanism. If it is being collected, it is likely in the form of spreadsheets or other PC-based applications and has to be integrated from there.

In summary, organizations that have reached the goal of optimizing a forecasting process, have done so by combining the ingredients of up-to-date forecast software systems, a people-driven process of capturing and integrating market events, and measuring / tracking forecast accuracy to obtain continuous improvement.